Skip to content

CS Masters Thesis. Scalable evolutionary Go using neural networks.

License

Notifications You must be signed in to change notification settings

wduncanfraser/scalable_go

Repository files navigation

Build Status

Scalable Go

Purpose

CS Masters Project. Scaling evolutionary neural networks for Go artificial intelligence.

Usage

Training

  • Once compiled, create a directory in the same location as the binary called "size<board size>set<set number>". For example, "size3set1".
  • Run training with ./scalable_go_training <board_size> <set> <start_generation> <end_generation> <uniform> <scaled>.
  • Uniform and scaled are booleans (enter 0 or 1) that determine if the network is uniform, and whether it is scaling up from a smaller network. If scaling up, "importnetworks.txt" much be present, which should be a copy of "lastbestnetworks.txt" from previous training on one size smaller board.

Comparison

  • Run comparison with ./scalable_go_comparison <board_size> <set1_name> <set1_uniform> <set2_name> <set2_uniform>. Example: ./scalable_go_comparison 5 size5set2 0 size5set6 1
  • set1_uniform and set2_uniform are booleans (enter 0 or 1) that determine if the network is uniform.

Benchmark

  • Run gogamenn benchmark with ./benchmark_gogamenn <board_size> <iterations>. Benchmark will return total time to complete iterations and iterations per second.

Structure

  • gogame/: Library for defining Go game, board, and move generation
  • neuralnet/: Neural Network Library
  • gogamenn/: Library defining NeuralNet wrapper for Go and helper functions.
  • gogameab/: Library defining AB Pruning algorithm.
  • tests/: Units and regression tests
  • benchmark_neuralnet.cpp: Basic benchmark of neural network performance.
  • benchmark_gogamenn.cpp: Basic benchmark of gogamenn performance.
  • benchmark_19x19ab_prune.cpp: Basic benchmark of worst case AB prune on 19x19 board with 0 ply.
  • scalable_go_comparison.cpp: Compares 2 sets of training results.
  • scalable_go_training.cpp: Training algorithm.

Neuralnet Structure

Layer 1 Subsection NeuralNet Node Counts

  • Input Layer: subsection size ^2, so 9 for a 3x3 subsection
  • HL1: Input Layer * (4/3) rounded down, so 12 for a 3x3 subsection
  • HL2: HL1 * (1/4) rounded down, so 3 for a 3x3 subsection
  • Output Layer: 1

Layer 2 Neuralnet Nodes

  • Input Layer: Outputs from all Layer 1 networks, in addition to pieces placed, friendly, and opponent prisoner counts
  • HL1: Input Layer * (2/3) rounded down
  • Output Layer: 1

External Resources

Legal

Author: W. Duncan Fraser

License: Apache License V2